Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
15
,
No.
1
,
Febr
uary
20
25
, pp.
728
~
740
IS
S
N:
20
88
-
8708
, DO
I: 10
.11
591/ij
ece.v
15
i
1
.
pp
728
-
740
728
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om
Enhan
cing PETRON
AS
share p
rice
forec
asts: a
hybri
d
H
olt
integrate
d movi
ng ave
rage
Nu
ri
n
Qistin
a M
oham
ad
F
oz
i
1
, N
ur
ha
s
niza Idha
m
Ab
u
Ha
s
an
1
,
A
zl
an Ab
dul
Aziz
2
,
Siti Meri
am
Z
ahari
3
,
Mog
ana
Da
rs
hini
G
ang
ga
y
ah
4
1
Co
lleg
e of
Co
m
p
u
tin
g
,
Info
rm
atics
an
d
M
ath
em
ati
cs,
Un
iv
ersiti T
ek
n
o
lo
g
i M
ARA
Perak B
ranch
,
Tapah
,
Mal
ay
sia
2
Co
lleg
e of Co
m
p
u
tin
g
,
Info
rm
atics
an
d
M
ath
em
ati
cs,
Un
iv
ersiti T
ek
n
o
lo
g
i M
ARA
Perlis B
ranch
,
Ar
au
,
M
ala
y
sia
3
Co
lleg
e of Co
m
p
u
tin
g
,
Info
rm
atics
an
d
Mathemati
cs,
Un
iv
ersiti T
ek
n
o
lo
g
i M
ARA
Selan
g
o
r
,
Sh
ah
Al
am
,
M
al
ay
sia
4
Sch
o
o
l of Bu
sin
es
s, M
o
n
ash
Univ
ers
ity
M
alay
sia,
Sela
n
g
o
r,
Malays
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
pr
4,
2024
Re
vised
Sep 4
,
2024
Accepte
d
Oct
1,
2024
Understa
nding
t
he
var
ia
t
ions
i
n
PETRONA
S
share
price
o
ver
t
im
e
is
im
porta
n
t
for
i
mprovi
ng
th
e
fo
rec
ast
accuracy
of
PETRONA
S
share
pri
ce
s
to
provide
stak
e
holde
rs
with
reli
abl
e
analyses
fo
r
future
ma
rk
et
pre
dictions.
The
ref
or
e,
the
ma
in
obj
ec
t
ive
of
th
is
stu
dy
is
to
im
prove
the
ac
cur
ac
y
of
PETRONA
S
share
pr
ice
by
u
ti
l
izing
a
hybrid
Hol
t
me
thod
wi
th
th
e
movi
ng
ave
rag
e
(MA
)
from
the
Box
-
Je
nkins
mode
l
.
Holt's
me
thod
wi
ll
addr
ess
li
ne
ar
tre
nds
for
non
-
stat
ion
ary
d
at
a
,
while
MA
will
an
al
y
ze
resid
ual
aspe
ct
s
of
the
data.
Th
is
com
b
ina
t
ion
tr
a
nsforms
non
-
stationary
d
at
a
in
to
stat
ionary
by
rem
oving
no
i
se
and
av
era
gin
g
out
fluctuation
s.
The
se
conda
r
y
dat
a
used
in
thi
s
study
co
nsists
of
dai
ly
o
bserva
ti
on
fro
m
b
ursa
Mala
ysia
,
the
offi
ci
a
l
nat
ion
al
stock
e
xcha
nge
of
Mal
ays
ia
,
cove
ring
the
per
iod
fro
m
Janua
ry
3
,
2000,
to
October
2,
2023
.
Th
e
s
tudy
en
com
pass
es
both
low
and
high
shar
e
pric
e
sc
ena
rios
.
The
models’
p
er
forma
nc
e
was
c
ompa
red
using
v
ari
ous
err
or
me
trics
ac
ross
diffe
ren
t
trainin
g
and
te
sting
sp
li
ts.
Th
e
findi
ng
s
hig
hl
ight
tha
t
th
e
propose
d
hybrid
[Hol
t
–
MA
]
mode
l
ca
l
l
ed
Hol
t
integra
t
ed
movi
ng
ave
rag
e
(
HIM
A
)
i
mprove
s
the
a
cc
ur
ac
y
of
f
ore
ca
st
ing
mod
el
with
th
e
smal
le
st
err
ors
f
or
both
d
ai
ly
lo
w
and
high
shar
e
pri
ce.
Th
e
HI
MA
mode
l
dem
onstra
te
s
si
gnifi
c
ant
pote
n
t
ia
l
,
p
ar
ticul
arl
y
in
red
u
ci
ng
re
sidual
s
and
im
proving
pre
di
ct
ion
accuracy.
Ke
yw
or
d
s
:
Accurac
y
Au
t
or
e
gr
e
ssive
integ
rated
movin
g
a
ver
a
ge
Dam
ped tre
nd
method
Ho
lt
met
hod
Time
ser
ie
s
f
oreca
sti
ng
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
BY
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
Nurhasniza
I
dham Ab
u Hasa
n
Coll
ege
of
C
omp
uting,
Infor
mati
cs an
d Ma
themat
ic
s,
Un
i
ver
sit
i Te
knologi
M
AR
A
Pe
ra
k
Bra
nc
h
35400, Ta
pa
h R
oad, Pe
rak D
aru
l R
idz
ua
n,
M
al
aysia
Emai
l:
nur
hasni
za@uitm.e
du.
my
1.
INTROD
U
CTION
Petrolia
m
Nasi
on
al
Be
rh
a
d,
know
n
as
P
ET
RON
AS
is
Ma
la
ys
ia
’s
fore
mo
st
inte
gr
at
e
d
oil
a
nd
gas
enter
pr
ise
a
nd
has
bec
ome
a
major
playe
r
in
the
global
energ
y
sect
or.
This
sect
or
op
erates
within
a
mil
ie
u
sh
a
ped
by
fl
uc
tuati
on
s
i
n
inte
rn
at
io
nal
oil
pri
ces,
ge
opolit
ic
al
com
plexiti
es,
te
ch
no
l
og
ic
al
ad
van
ce
me
nts
,
an
d
su
sta
ina
bili
ty
impe
rati
ves
[
1]
.
The
s
har
e
pr
i
ce
of
PE
TRO
NAS
enca
psula
te
s
the
cum
ulati
ve
impact
of
thes
e
div
e
rs
e
facto
rs
,
se
rv
i
ng
as
a
r
eflect
ion
not
only
of
the
c
ompan
y’s
fina
nci
al
pe
rformance
but
al
so
of
broad
e
r
tren
ds
within
the
e
nergy
sect
or
.
Give
n
it
s
f
unda
me
ntal
r
ole
i
n
unde
rp
i
nn
i
ng
the
natio
n’
s
e
conomic
fr
ame
w
ork,
t
he
vo
la
ti
li
ty
of
P
ETRO
NA
S
’
s
ha
re
pri
ce
hold
s
sign
ific
a
nt
ra
mific
at
ion
s
for
M
al
aysia
’s
fi
na
ncial
mar
kets
an
d
broa
der
eco
nom
ic
sta
bili
ty
[2]
.
The
inte
rp
la
y
of
t
hese
fa
ct
ors,
c
ombine
d
with
the
c
omp
any’s
strat
egic
init
ia
ti
ves
an
d
fi
nanci
al
per
f
orman
ce,
co
ntrib
utes
to
the
intric
at
e
mo
sai
c
of
it
s
market
valuati
on
[3]
.
As
M
al
ay
sia
posit
ion
s
it
sel
f
as
a
reg
i
on
al
e
nerg
y
hub,
PE
TRON
AS
sha
r
e
pri
ce
is
no
t
on
l
y
ref
le
ct
ive
of
it
s
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
En
hancin
g
P
E
TRO
NA
S
s
hare
p
ric
e f
or
eca
sts:
a hyb
ri
d
H
olt i
ntegr
ated
…
(
Nuri
n
Qisti
na
Mo
hamad Fo
zi
)
729
performa
nce
but
is
al
so
em
blemat
ic
of
the
broa
der
tre
nd
s
a
nd
chall
en
ges
f
aci
ng
t
he
e
nerg
y
sect
or
i
n
the
Asia
-
Paci
fic
r
egi
on
[
4]
.
Acc
ur
at
e
f
or
eca
sti
ng
of
the
s
har
e
p
ric
e
of
PETR
ONAS
is
a
c
omplex
unde
rtaki
ng
tha
t
requires
a
nua
nced
underst
and
i
ng
of
t
he
uniq
ue
dyna
mics
at
pla
y
withi
n
t
he
e
nerg
y
s
ect
or
,
as
well
as
the
broa
der
ec
onomi
c
co
ntext
in
M
al
aysia
[5],
[6]
.
T
he
ob
je
ct
ive
of
t
his
st
udy
is
unra
velli
ng
the
ti
me
series
patte
rn
s
in
her
e
nt
in
PETRO
NAS
s
har
e
pr
ic
e,
f
or
ecast
f
uture
tre
n
ds
t
hro
ugh
ti
me
s
eries
te
ch
niqu
es,
an
d
crit
ic
al
ly asses
s the acc
ur
a
cy
of these
f
or
eca
sti
ng
meth
odol
og
ie
s
.
Fo
r
ecast
ing
th
e
PETR
ONAS
sh
a
re
pri
ce
ha
s
bec
om
e
inc
reasin
gly
imp
ort
ant
for
e
nerg
y
plan
ning
.
The
pr
ima
r
y
goal
s
incl
ud
e
e
sta
blishin
g
a
ppr
opriat
e
pri
ci
ng
an
d
ta
xatio
n
f
rame
w
orks,
assist
ing
i
n
pote
ntial
inv
est
me
nts
a
nd
de
ci
sion
-
m
akin
g
reg
a
r
ding
oil
rese
r
ves
to
e
nh
a
nce
e
ne
rgy
s
ecu
rity,
facil
it
at
ing
the
early
reso
l
ution
of
emissi
on
a
nd
con
ta
mi
natio
n
issues
,
e
na
bling
the
prov
isi
on
of
f
uture
e
nerg
y
dema
nd
s,
a
nd
identif
ying
national
in
fr
a
struc
ture
a
nd
e
nerg
y
re
quireme
nts
.
Acc
urat
e
f
or
e
cast
ing
s
uppo
rts
strat
egic
plan
nin
g
and
po
li
cy
-
ma
king,
en
surin
g
a
sta
ble
a
nd
secur
e
e
nerg
y
f
uture
[7]
.
G
iven
the
imp
ort
ance
of
oil
t
o
t
he
econom
y,
pr
oject
ing
P
ETRO
NAS
s
har
e
pri
ces
has
recei
ve
d
a
l
ot
of
f
oc
us
i
n
the
li
te
ratur
e.
The
decisi
on
on
wh
e
n
to
dri
ll
and
how
m
uch
to
he
dg
e
is
bas
ed
on
forecast
ing,
w
hich
is
in
her
e
ntl
y
an
im
perfect
sci
e
nce
.
T
hi
s
is
especial
ly
tr
ue
f
or
the
oil
industr
y,
a
glob
al
sect
or
c
hara
ct
erized
by
wi
dely
inc
onsist
ent
data
[8]
.
I
na
ccur
at
e
forecast
s
with
up
si
de
or
dow
ns
ide
bias
ca
n
aff
ect
pr
of
it
lo
ss,
with
dow
nsi
de
bias
ca
us
in
g
hi
gh
e
r
los
ses
du
e
t
o
lowe
r
pr
e
dicte
d
pr
ic
es
[
9]
.
P
ast
f
or
ecast
in
g
ap
proac
hes
ar
e
cr
ucial
f
or
unde
rstan
ding
why
petr
oleu
m
s
har
e
pr
ic
es
a
r
e
dif
fi
cult
to
pr
e
dict
[10]
.
The
refo
re,
a
stu
dy
on
forecas
ti
ng
P
ETRO
NA
S
sha
re
pri
ce
has
become
essenti
al
.
Ther
e
are
ma
ny
met
hods
f
or
in
vestigat
in
g
the
PE
TRO
NAS
s
har
e
pr
ic
e.
The
m
os
t
com
m
on
forecast
in
g
model
is
the
tra
diti
on
al
Ho
lt
me
thod
[
11]
.
M
uc
hay
a
n
[
12]
use
d
the
tradit
i
on
al
H
olt
m
et
hod
t
o
form
ulate
a
f
oreca
sti
ng
mod
el
,
determi
ning
w
hich
met
hod
pro
vid
es
m
or
e
acc
ur
at
e
predict
ion
s
for
ne
t
asset
value
(
N
AV)
pri
ce
m
oveme
nt
s
f
r
om
Ja
nu
a
r
y
1,
20
19,
t
o
Ja
nu
a
r
y
1,
2020.
The
stu
dy
f
ound
t
hat
t
he
tra
diti
on
al
Ho
lt
meth
od
ou
t
performe
d
the
oth
e
r
ben
c
hm
a
rk
models
,
w
hich
pro
duced
the
le
ast
error.
A
nother
study
cond
ucted
by
Bog
a
r
a
nd
G
ungo
r
[
13]
f
ound
that
t
he
tra
diti
on
al
Ho
lt
met
hod
was
superi
or
t
o
oth
e
r
m
odel
s
f
or
forecast
in
g
the
quantit
y
of
wa
ste
mobil
e
phones (
W
M
P
)
in
Tu
r
ke
y.
T
he
st
udy
fou
nd
that t
he
H
olt
meth
od
ca
n
accuratel
y
pre
dict
fu
t
ur
e
dat
a
po
i
nts
w
her
e
the
tren
d
is
e
it
her
inc
reasin
g
or
decr
easi
ng
at
a
co
ns
ta
nt
rate.
Fu
rt
hermo
re,
S
hukor
et
al.
[14]
disc
ov
e
red
that
the
tra
diti
onal
H
olt
meth
od
is
the
mo
st
accurate
forec
ast
in
g
method
f
or
the
stoc
k
mar
ket
pr
ic
es
of
c
r
ud
e
oil
a
nd
plati
num,
util
iz
ing
monthly
data
f
rom
Ja
nuar
y
2000
to
Decem
ber
20
16.
It
is
e
vid
e
nt
that
the
tra
diti
on
al
H
olt
m
et
hod
re
mains
a
r
el
evan
t
a
nd
wi
dely
us
e
d
fore
cast
ing
method
t
od
a
y.
Howe
ver,
due
to
the
po
te
ntial
li
mit
at
ion
of
the
tra
diti
on
al
Ho
lt
met
hod,
s
uch
as
se
ns
it
iv
it
y
to
par
a
mete
r
sel
e
ct
ion
“
α
a
nd
β
”
,
the
ass
umpti
on
of
a
li
nea
r
tren
d
in
ti
me
s
eries
fall
s
shor
t
wh
e
n
deali
ng
with
data
that
e
xh
i
bits
stron
g
sea
so
na
l
patte
r
ns
,
te
nd
to
overe
sti
mate
or
unde
resti
mate
,
es
pe
ci
al
ly
in
lo
ng
-
te
r
m
forecast
s a
nd hand
li
ng
of str
uc
tural c
hanges
[15],
[16],
[17]
.
Re
cognizi
ng
t
he
se
s
hortco
min
gs
,
Ga
r
dn
e
r
a
nd
M
c
Kenzi
e
[
18]
intr
oduce
d
the
dam
pe
d
t
re
nd
met
hod
(D
T
M)
in
1985,
wh
ic
h
inclu
des
a
dam
pen
i
ng
facto
r
to
re
du
ce
t
he
im
pa
ct
of
tre
nds
ov
er
ti
me.
H
owe
ver,
this
method
sti
ll
strug
gles
t
o
outper
form
oth
e
r
smoothi
ng
met
hods
c
on
sist
e
nt
ly
a
nd
sti
ll
fal
ls
sho
rt
of
ac
hi
eving
op
ti
mal
f
or
eca
st
accu
racy
[19]
.
T
he
refor
e
,
this
stu
dy
is
c
onduct
ed
to
imp
rove
the
ac
cur
ac
y
pe
rformance
o
f
PETRO
NAS
s
har
e
pri
ces
by
com
bin
in
g
the
tradit
ion
al
H
olt
meth
od
with
the
m
ovin
g
a
ve
rag
e
(
MA)
f
r
om
t
he
Box
-
Je
nk
i
ns
m
od
el
.
T
he
H
olt
meth
od
ef
fecti
vely
ca
ptures
l
inear
tre
nds
a
nd
handle
s
tre
nds
for
non
-
sta
ti
on
a
r
y
data.
T
he
pu
rpose
of
MA
in
t
his
hy
br
id
model
is
the
capa
bili
ty
of
MA
t
o
anal
yze
re
sidu
al
s
t
hat
exis
t
in
the
data
bein
g
st
udie
d.
Be
sides
t
hat,
MA
helps
co
rr
ect
er
rors
by
a
dju
sti
ng
pa
st
f
or
ecast
er
r
or
s
a
nd,
at
t
he
same
ti
me,
imp
ro
ve
the
per
ce
nta
ge
acc
ur
ac
y
of
m
odel
pre
dicti
on
s
[20
],
[
21]
.
This
co
mb
inati
on
tra
ns
f
orms
non
-
sta
ti
on
a
ry
data
int
o
sta
ti
onar
y
data
by
r
emo
ving
noise
an
d
a
ver
a
ging
out
fl
uctuati
ons
[22
]
.
Othe
r
than
that, the
oth
er
pur
po
se
of this
study
is t
o
c
on
tribu
te
to
t
he
e
xisti
ng
body
of k
no
wled
ge,
prov
i
ding stake
holde
rs
with
a
rob
us
t
analyti
cal
f
oundat
ion
f
or
a
ntici
pating
fu
t
ur
e
mar
ket
m
ov
e
ments.
The
i
nsi
gh
ts
gen
e
rate
d
f
ro
m
this
resea
rch
ar
e
ex
pected
t
o
e
mpowe
r
in
vest
or
s
,
a
nalysts,
a
nd
poli
cyma
ke
rs
with
the
knowle
dge
necess
ary
t
o
nav
i
gate the
cha
ll
eng
es a
nd
oppo
rtu
niti
es p
r
esented
by t
he e
vo
l
ving e
nerg
y
la
ndsca
pe
i
n Malaysia.
T
h
i
s
r
es
e
ar
ch
a
r
t
i
c
l
e
i
s
s
tru
ctu
re
d
a
s
f
o
l
low
s
:
s
ec
t
i
on
1
pr
ov
id
e
s
th
e
ba
c
kg
ro
un
d
of
t
h
e
s
tud
y
,
i
t
s
ob
j
e
c
t
i
ve
s
,
and
an
ov
e
rv
i
ew
of
t
r
ad
i
t
io
na
l
m
e
t
ho
d
s
a
l
on
g
s
id
e
th
e
hy
b
ri
d
H
o
l
t
in
t
eg
ra
t
e
d
m
ov
ing
av
er
age
(
HIMA
)
mo
d
el.
S
e
c
t
io
n
2
re
v
i
ew
s
p
re
v
iou
s
wo
rk
s
by
o
th
e
r
r
e
se
a
rc
he
r
s
in
th
e
f
i
e
ld
.
S
e
c
t
io
n
3
of
fe
rs
a
d
e
t
a
i
l
ed
d
e
s
cr
ip
t
i
on
of
th
e
Ho
l
t
m
e
tho
d
,
DT
M
,
and
a
u
tore
gr
e
ss
i
v
e
mo
vi
ng
a
ve
r
age
(
AR
IM
A
)
m
od
e
l,
a
lon
g
w
i
t
h
th
e
con
s
tru
c
t
io
n
pr
o
c
e
ss
o
f
th
e
HI
M
A
m
od
e
l.
S
ec
t
i
on
4
d
is
cu
s
s
es
t
he
r
e
su
l
t
s,
an
d
s
e
c
t
i
on
5
c
on
c
lud
e
s
t
h
e
s
t
ud
y.
2.
LIT
ERATUR
E REVIE
W
M
ost
of
the
res
earche
rs
s
ugge
st
a
hybri
d
model
that
pr
oduc
es
more
rob
us
t
an
d
en
ha
nced
forecast
in
g
accurac
y
[
23]
.
Be
sides,
S
hetty
an
d
Is
mail
[
24]
in
dicat
ed
th
at
the
pe
rfo
rm
ance
of
hy
br
id
forecast
in
g
models
i
s
bette
r
tha
n
an
y
tra
diti
onal
or
ben
c
hm
a
r
k
models.
A
l
ot
of
w
ork
ha
s
been
c
onduct
ed
re
ga
rd
i
ng
the
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
728
-
740
730
com
bin
at
io
n
of
the
tra
diti
on
al
m
odel
with
ot
her
m
odel
s,
s
uc
h
as
ne
ur
al
ne
tworks
,
w
hich
can
be
fou
nd
i
n
the
li
te
ratur
e.
T
he
ulti
mate
fo
c
us
of
the
de
velo
pme
nt
of
this
hybr
i
d
m
odel
i
s
t
o
yield
a
more
accurate
f
or
eca
sti
ng
model
with
the
le
ast
e
rror
perce
ntage.
Seve
r
al
hy
br
id
mod
el
s
ha
ve
been
i
mp
le
me
nted
in
pre
vious
stu
di
es
f
or
Ho
lt
met
hod.
Haque
et
al.
[
25]
propose
d
a
modifie
d
H
olt’s
meth
od
f
or
sh
ort
-
te
r
m
f
ore
cast
ing
m
odifi
cat
ion
that
us
es
ti
me
series
data
fro
m
1991
t
o
20
13
an
d
forecast
s
for
fi
ve
c
on
s
ecuti
ve
year
s
f
rom
20
14
t
o
2018.
It
inco
rpor
at
es
re
cent
tren
d
valu
es
as
a
weig
ht
par
a
mete
r
a
nd
pr
e
vious
tre
nd
s.
As
a
res
ult,
the
m
od
i
fied
m
et
hod
yields
f
oreca
st
values
cl
os
er
to
act
ual
val
ue
s,
with
f
or
ec
ast
values
sli
ghtl
y
hi
gher
a
nd
m
or
e
al
ig
ne
d
with
ob
s
er
ved
data.
The
stu
dy
c
oncl
ud
es
t
he
ne
w
hybri
d
met
hod
pr
ov
i
des
a
m
or
e
acc
urat
e
an
d
reli
able
est
i
mate
of
real
values
.
Moh
a
mme
d
[
26]
s
uggested
a
ne
w
hybri
d
m
od
el
f
or
ti
me
se
ri
es
f
oreca
sti
ng
cal
le
d
AR
-
H
ol
t
(
p+5)
to
en
ha
nce
predict
io
n
acc
uracy
an
d
rob
us
tness
.
T
his
m
odel
c
ombines
the
H
ol
t
method
wi
th
the
auto
regressive
(A
R
)
m
odel
.
The
perf
or
ma
nce
of
t
he
AR
-
H
olt
(
p+5)
m
od
el
is
bette
r
than
oth
e
r
me
thods,
wh
ic
h
Y
ule
-
Walke
r,
Bu
rg,
resi
du
al
aut
o
-
co
var
ia
nce
(
RA),
le
ast
s
quares
,
modifie
d
c
ovaria
nce,
and
le
ast
median
s
qu
a
re
s
(L
M
S
) met
ho
ds
.
M
ore
ov
e
r,
Eg
r
iog
lu
an
d
Ba
ş
[
27]
int
rod
uced
a
modifie
d
H
ol
t’s
li
near
tre
nd method b
ase
d
on
pa
rtic
le
swarm
opti
miz
at
ion
(
PS
O)
to
impro
ve
f
or
ec
ast
ing
acc
ur
ac
y.
T
he
stu
dy
use
d
sim
ulate
d
data
set
s
with
li
near
and
quad
rati
c
tren
ds
,
a
s
well
as
real
-
world
da
ta
fr
om
the
Is
ta
nbul
stoc
k
ex
change
,
c
onsist
ing
of
t
wo
dif
f
eren
t
ti
me
series
obs
erv
e
d
dail
y
bet
ween
Februa
r
y
1,
20
09
t
o
M
a
y
29,
2009
,
a
nd
A
pr
il
1,
2010
to
M
ay
31,
2010.
The
meth
od
updates
tre
nds
and
values
a
nd
i
ncor
porates
sec
ond
-
orde
r
update
f
ormul
as.
C
onseq
ue
nt
ly,
t
he
study
co
nclu
de
s
that
the
pr
opos
e
d
meth
od
outpe
rforms
t
he
tra
diti
on
al
H
ol
t’s
meth
od
in
te
rms
of
forec
ast
ing
performa
nce.
Be
sides,
Li
u
and
W
u
[28
]
introd
uced
a
modifie
d
H
olt’s
e
xponentia
l
smoot
hing
(
M
H
ES)
method
for
pr
e
dicti
ng
hous
i
ng
pri
ces.
This
method
a
dju
sts
histo
rical
data
weig
hts
a
nd
s
moothin
g
par
a
mete
rs
base
d
on
sam
pl
e size
. Housin
g pr
ic
e
data f
rom Ku
nm
in
g, Chan
gc
hun, X
uz
hou, an
d Ha
ndan
s
panni
ng
Ja
nu
a
r
y
2015
to
Aug
us
t
2019
were
util
iz
ed.
Co
mp
a
rati
ve
ana
lysis
re
vealed
that
the
ne
wly
dev
el
op
e
d
w
hal
e
o
ptimi
zat
io
n
al
gorithm
-
m
od
ifie
d
H
olt’s
e
xpone
ntial
sm
oothin
g
(
W
OA
-
M
H
ES
)
met
ho
d
e
xhibit
ed
s
uperi
or
performa
nce
over
co
nventio
na
l
models
“
bac
kpr
op
a
gatio
n
ne
ur
al
netw
ork
(
BPN
N
)
,
(gre
y
model
(GM
)
(
1,1)),
and
AR
IMA
,
”
dem
onstrat
in
g
re
duce
d
pre
dicti
on
e
rro
rs
an
d
fa
ste
r
c
ompu
ta
ti
on
ti
mes.
T
hese
fi
nd
i
ngs
highli
gh
t
the
a
pp
li
cabil
it
y
of
the
ne
w
hybr
id
m
odel
in
f
oreca
sti
ng
f
or
hous
i
ng
pr
ic
e
mar
ket
in
vesto
rs
a
nd
po
li
cy
mak
ers
.
Fu
rt
hermo
re,
s
om
e
of
the
pr
evio
us
st
ud
ie
s
discusse
d
hybr
id
m
od
el
s
f
or
ARI
M
A
.
T
he
researc
h
by
R
avicha
ndran
et
al.
[29]
ai
m
ed
to
f
or
eca
st
the
yield
a
nd
pro
du
ct
ivit
y
of
f
ood
gr
ai
ns
w
it
hin
the
ag
ricult
ur
al
sect
or
thr
ough
the
a
pp
li
cat
ion
of
ti
me
-
series
analysis.
T
he
da
ta
set
in
that
study
e
nc
ompas
ses
ti
me
se
ries
data
on
the p
r
oducti
on
an
d
yield o
f
var
i
ous
oilsee
d
cr
op
s
s
pa
nn
i
ng
f
rom
19
50
-
51
to 2
01
5
-
16. Th
e
st
udy
e
mpl
oy
s
a
hybri
d
meth
od
ology
th
at
inte
gr
at
es
a
uto
re
gr
essive
movi
ng
ave
ra
ge
(
ARI
M
A
)
m
od
el
s
with
a
rtific
ia
l
neural
netw
orks
(
A
N
N)
as
(
ARI
MA
-
ANN
)
mode
l.
T
he
fin
dings
de
m
on
st
rate
t
hat
t
he
ARI
MA
-
ANN
hybr
i
d
model
ou
t
performs
t
he
ind
ivi
du
al
A
RIMA
an
d
ANN
m
od
el
s
in
te
rms
of
acc
ur
ac
y
an
d
ef
fecti
ve
ness.
Ec
he
varri
a
and
Ar
a
nas
[
30]
e
mp
lo
ye
d
c
ons
um
e
r
pr
ic
e
i
ndex
(CPI)
data
from
t
he
P
hili
pp
i
nes
a
nd
it
s
re
gions
for
t
he
yea
r
2022,
util
iz
ing
a
hy
br
id
ARI
M
A
-
ANN
a
pp
ro
ac
h
t
o
e
nhan
ce
CPI
f
or
e
cas
ti
ng
acc
uracy
.
As
a
re
su
lt
,
the
hy
br
id
ARI
M
A
-
ANN
m
od
el
s
c
on
si
ste
ntly
s
urpas
sed
sta
ndal
on
e
AR
IMA
m
odel
s,
delive
rin
g
more
preci
s
e
an
d
reli
able
f
or
eca
sts
ov
e
r
a
n
ext
end
e
d
per
i
od.
Ov
e
rall
,
the
re
su
lt
s
of
the
pr
e
vious
stu
dies
s
howe
d
that
the
hybri
d
model was a
ble to in
crease t
he
p
re
dicti
on’s ac
cur
ac
y
in
ma
ny
a
ppli
cat
ion
s
, m
aki
ng
it
a potenti
al
can
dida
te
f
or
ti
me series fo
r
ecast
ing
a
naly
s
is.
3.
METHO
D
This
researc
h
ai
ms
to
co
mp
a
re
the
perform
ance
of
the
tra
diti
on
al
H
olt
method,
DT
M
,
ARI
MA,
an
d
a
ne
w
hybri
d
tradit
ion
al
H
ol
t
an
d
m
ovin
g
aver
a
ge
cal
le
d
H
IMA
m
od
el
em
ployin
g
ti
me
-
se
ries
a
nalysis
to
forecast
the
lo
w
a
nd
high
P
E
TRON
AS
s
hare
pri
ces.
Data
analysis
was
pe
rformed
usi
ng
R
pr
ogram
ming
a
nd
Excel.
Fig
ur
e
1 presents
a sc
he
mati
c
represe
ntati
on
of the
re
search
f
l
ow of this
stu
dy.
The
resea
rch
method
ology
c
omprises
fi
ve
disti
nct
sta
ges
.
It
beg
i
ns
with
ex
plorat
ory
da
ta
analysis
(EDA),
a
crit
ic
al
phase
for
su
m
marizi
ng
t
he
pr
ima
r
y
fea
tures
of
t
he
ti
me
se
ries
data.
T
his
sta
ge
in
vo
l
ve
s
d
et
ect
ing
tre
nd
patte
r
ns
,
outl
ie
rs,
an
d
mis
sing
values
i
n
the
dataset
[
31]
.
T
he
data
is
then
cl
ean
ed,
a
nd
incomple
te
e
ntries
are
a
ddres
sed
by
re
movi
ng
missi
ng
va
lues
[
32]
.
S
ubseq
uen
tl
y,
the
com
plete
datas
et
is
div
ide
d
int
o
two
s
ubset
s:
trai
nin
g
a
nd
te
sti
ng
.
T
his
pa
r
ti
ti
on
ing
is
im
portant
f
or
as
sessing
t
he
m
od
el
's
pe
r
forma
nce.
The
par
ti
ti
on
i
ng
strat
eg
y
rel
ie
s
on
ti
me
se
ries
cr
os
s
-
vali
dation
to
e
nha
nce
the
accu
ra
cy
of
forecast
values
, as
il
lustrate
d
i
n
Fi
gure
2.
Figure
2
pres
ents
the
pa
rtit
ion
in
g
of
a
f
ull
dataset
i
nto
five
disti
nct
set
s,
eac
h
w
it
h
va
ry
i
ng
pro
portions
of
trai
ning
a
nd
t
est
ing
s
ubset
s.
In
s
et
1,
99%
of
t
he
data
is
al
locat
ed
f
or
trai
ning
an
d
1%
for
te
sti
ng
,
w
hile
i
n
s
et
2,
95%
is
us
e
d
f
or
trai
nin
g
a
nd
5%
f
or
te
sti
ng
.
Set
3
a
ssign
s
90%
t
o
trai
ning
a
nd
10
%
to
te
sti
ng
,
s
et
4
f
ollows
with
80%
f
or
t
rainin
g
an
d
20%
f
or
te
sti
ng
,
a
nd
s
et
5
al
locat
es
70
%
of
the
da
ta
to
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
En
hancin
g
P
E
TRO
NA
S
s
hare
p
ric
e f
or
eca
sts:
a hyb
ri
d
H
olt i
ntegr
ated
…
(
Nuri
n
Qisti
na
Mo
hamad Fo
zi
)
731
trai
ning
a
nd
30%
t
o
te
sti
ng.
This
pro
gress
ive
par
ti
ti
on
i
ng
a
ppr
oach
s
uppo
rts
cr
os
s
-
va
li
dation
by
al
lowing
model
e
valuati
on
ac
ro
s
s
dif
fe
ren
t
trai
ni
ng
a
nd
te
sti
ng
rati
os
[
33]
,
[34
],
[
35]
.
Ta
ble
1
pr
e
sents
t
he
distri
bu
ti
on
su
m
mar
y
of
th
e
trai
ning
an
d
t
est
ing
set
s
f
or
daily
hi
gh
an
d
low
P
ETRO
N
AS
s
har
e
pri
ces
across
fi
ve
di
sti
nct
set
s.
Figure
1. Re
se
arch
f
rame
wor
k of t
his st
udy
Figure
2.
Data
sp
li
t of ti
me se
ries
cr
os
s
-
valid
at
ion
The
dataset
is
div
ide
d
i
nto
t
r
ai
nin
g
an
d
te
st
ing
set
s
acr
os
s
five
co
nf
i
gura
ti
on
s,
each
va
r
ying
in
the
pro
portion
of
data
al
locat
e
d
for
trai
ning
a
nd
te
sti
ng.
T
he
se
co
nfi
gurati
ons
dif
fer
in
te
r
ms
of
du
rati
on
an
d
Do
es th
e tr
en
d
exi
st in
th
e PE
TRO
NAS
sh
are
p
rice
d
ata?
Start
Exp
lo
ratory
d
ata a
n
aly
sis
(E
DA
)
Mod
el Develo
p
m
e
n
t
Mod
el E
v
alu
atio
n
End
Data Pa
rtition
in
g
Tr
ain
in
g
Testin
g
Mod
el Per
form
an
ce
No
Yes
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
728
-
740
732
sample
siz
e,
de
sign
e
d
t
o
asses
s
the r
ob
us
tnes
s
an
d
reli
abili
ty o
f
t
he
pr
e
dict
ive
m
od
el
acr
oss
di
ver
se
sce
na
rios.
The
sec
onda
ry
data
util
iz
ed
in
this
stu
dy
c
omprises
daily
obser
vatio
ns
from
bursa
Ma
la
ys
ia
,
the
na
ti
on
al
stock
e
xc
hang
e
of
M
al
ay
sia
,
sp
an
ni
ng
t
he
per
i
od
from
Ja
nu
a
r
y
3,
2000
to
Octo
be
r
2,
2
02
3.
T
he
a
na
lysis
include
s
both
l
ow
a
nd
hi
gh
-
s
har
e p
rice
sce
na
rios.
Out
of
th
e
or
i
gin
al
5,9
43
dataset
s,
tw
o
obser
vations
c
on
ta
in
missi
ng
val
ues
,
res
ulti
ng
in
a
final
t
otal
of
5,9
41
dataset
s.
This
final
dataset
will
be
div
i
ded
acco
r
ding
to
th
e
five
diff
e
re
n
t
config
ur
at
io
n
set
s
as
il
lustrate
d
in
Ta
ble
1.
The
ne
xt
ph
ase
involves
t
he
de
velo
pme
nt
of
a
hybri
d
H
olt
a
nd
movin
g
a
verage
(
M
A
)
m
odel
.
I
n
this
phas
e,
five
data
c
onfi
gurati
ons
a
r
e
em
ploye
d
t
o
assess
the
acc
ur
ac
y
of
PE
TRO
NAS
sh
a
re
pr
ic
e
pr
edict
ion
s
by
c
ompa
ri
ng
the
t
rad
it
io
nal
m
odel
with
t
he
pro
po
s
ed
model. Fi
gure
3
il
lustrate
s t
he
model
dev
el
opme
nt pr
ocess of
the
HIM
A mo
del.
Table
1
.
T
est
in
g
a
nd
trai
ni
ng
set
f
or
daily
hi
gh and l
ow
Set
Data
p
artition
Percentag
e
(%)
Du
ration
Sam
p
le
s
iz
e
Set 1
Tr
ain
in
g
99
Jan
u
ary 3
,
2
0
0
0
–
Ju
ly
4, 20
2
3
5
,88
2
Testin
g
1
Ju
ly
5, 20
2
3
–
Oct
o
b
er
2
,
2
0
2
3
61
Set 2
Tr
ain
in
g
95
Jan
u
ary 3
,
2
0
0
0
–
Ju
ly
8, 20
2
2
5
,64
4
Testin
g
5
Ju
ly
12
,
2
0
2
2
–
Octo
b
er
2
,
2
0
2
3
299
Set 3
Tr
ain
in
g
90
Jan
u
ary 3
,
2
0
0
0
–
Ap
ril
2
0
,
2
0
2
1
5
,34
7
Testin
g
10
Ap
ril
2
1
,
2
0
2
1
–
Octo
b
er
2
,
2
0
2
3
596
Set 4
Tr
ain
in
g
80
Jan
u
ary 3
,
2
0
0
0
–
No
v
em
b
er
2
1
,
2
0
1
8
4
,75
3
Testin
g
20
No
v
em
b
er
2
2
,
2
0
1
8
–
Octo
b
er
2
,
2
0
2
3
1
,19
0
Set 5
Tr
ain
in
g
70
Jan
u
ary 3
,
2
0
0
0
–
Ju
ly
1, 20
1
6
4
,15
9
Testin
g
30
Ju
ly
4,
2016
–
Oct
o
b
er
2
,
2
0
2
3
1
,78
4
Figure
3. Mo
de
l
dev
el
opme
nt of
H
I
M
A
The
pr
opos
e
d
model
de
velo
pme
nt
sta
rts
wit
h
est
imat
ing
t
wo
e
quat
ion
s
,
wh
ic
h
are
le
ve
l
an
d
tre
nd
as
sta
te
d
in
(
2)
a
nd
(
3),
res
pecti
ve
ly.
The
n,
t
he
forecast
val
ue
in
(1)
will
be
ge
ner
at
e
d.
T
he
Ho
lt
resi
du
al
va
lues,
wh
ic
h
represe
nt
the
dif
fer
e
nc
es
bet
ween
the
act
ua
l
data
an
d
th
e
f
orec
ast
ed
values
,
a
re
al
so
cal
cula
te
d
as
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
En
hancin
g
P
E
TRO
NA
S
s
hare
p
ric
e f
or
eca
sts:
a hyb
ri
d
H
olt i
ntegr
ated
…
(
Nuri
n
Qisti
na
Mo
hamad Fo
zi
)
733
represe
nted
in
(12).
Lat
er
,
th
e
Ho
lt
resid
ual
will
be
us
ed
in
the
MA
proc
ess.
T
o
e
ns
ure
the
resi
du
al
c
an
be
us
e
d
f
or
t
he
MA
pr
ocess,
t
he
assumpti
on
of
the
sta
ti
on
a
ry
needs
to
be
c
he
cked.
A
sta
ti
onar
y
se
ries
is
a
series
that
fluct
uates
ra
ndom
l
y
a
round
some
fix
ed
values
(
me
an
var
ia
nce
or
a
ny
c
on
sta
nt
val
ue)
[
36]
.
If
t
he
assumpti
on
is
vio
la
te
d,
a
dif
f
eren
ci
ng
pr
oce
ss
will
be
a
pp
l
ie
d
[
37]
.
T
he
M
A
m
odel
is
est
imat
ed
usi
ng
th
e
Ho
lt
re
sid
uals
as
input.
The
M
A
f
or
eca
st
va
lues
are
the
n
ge
ner
at
e
d
as
in
(
11).
Finall
y,
t
he
hy
br
i
d
model
com
bin
es
the
Ho
lt
a
nd
M
A
f
or
ecast
s
to
pro
du
ce
the
final
ou
t
pu
t,
ei
ther
by
ad
ding
or
s
ub
t
racti
ng
to
pro
duce
the f
i
nal foreca
ste
d values
[
H
olt ±
MA]
as i
n (13).
3.1.
Tr
adi
tional
m
et
hods
The
i
mp
le
me
nt
at
ion
a
nd
e
valuati
on
of
f
or
ec
ast
ing
te
c
hniq
ues,
inclu
di
ng
the
tra
diti
on
al
or
existi
ng
methods
s
uc
h
as
the
H
olt
method,
DT
M
,
and
ARI
MA
model.
F
ur
th
e
r
detai
ls
are
di
scusse
d
in
th
e
nex
t
su
bse
ct
ions
3.1.1. Hol
t
m
e
th
od
Ho
lt
[
38],
[39]
de
vel
op
e
d
the
H
olt
meth
od
i
n
1957
,
w
hich
enh
a
nce
d
simp
le
ex
pone
ntial
smoothi
ng
to pre
dict data
with a t
rend.
Fo
r
ecast
:
+
=
+
×
(1)
Level:
=
+
(
1
−
)
(
−
1
+
−
1
)
(2)
Trend:
=
(
−
−
1
)
+
(
1
−
)
−
1
(3)
wh
e
re
de
note
s an
e
sti
mate
of
the level
of the
ser
ie
s at ti
me
,
de
note
s a
n
est
imat
e o
f
the t
re
nd
“
slo
pe
”
of
the series
at ti
me
. Two
s
moothin
g
c
onsta
nt
s,
an
d
, w
it
h v
al
ues betwee
n 0 a
nd 1, a
re
use
d
in
this
meth
od,
is t
he
s
moothi
ng p
a
ramete
r
f
or
the le
vel, a
nd
is t
he
s
moot
hing
par
a
mete
r
for
t
he
tre
nd.
3.1.2. D
amp
ed
tren
d me
thod
(D
T
M
)
Gard
ner
a
nd
M
c
ken
zi
e
[
39]
intr
oduce
d
the
D
TM,
w
hich
eff
ect
ivel
y
re
duces
the
im
pac
t
of
t
he
tren
d,
causin
g
it
to
le
vel
off
at
a
s
pe
ci
fic
po
i
nt
in
t
he
f
uture.
T
his ap
pr
oach
c
onta
ins
a
da
mp
i
ng f
act
or
to
mit
ig
at
e
the
influ
e
nce
of
ol
der
data.
I
n
a
ddit
ion
t
o
th
e
and
par
a
mete
rs
,
the
re
is
a
da
mp
in
g
pa
ramet
er
(
)
t
hat
va
r
ie
s
from 0 t
o 1.
Fo
r
ecast
:
+
=
+
(
+
2
+
⋯
+
)
(4)
Level:
=
+
(
1
−
)
(
−
1
+
−
1
)
(5)
Trend:
=
(
−
−
1
)
+
(
1
−
)
−
1
(6)
3.1.3
Au
t
ore
gr
essive
in
teg
r
at
e
d m
ov
in
g a
verage
(
AR
I
MA)
ARI
M
A
m
odel
s
fi
nd
a
pp
li
cat
ion
s
in
the
pr
e
di
ct
ion
of
ti
me
series
data,
w
hi
ch
e
ncompa
ss
seq
ue
ntial
data
points
gather
e
d
or
do
c
ume
nted
at
regu
la
r
ti
me
interv
al
s.
These
models
co
mprise
three
pr
ima
ry
models:
a
uto
r
eg
ressive
(A
R)
,
i
nteg
rate
d
(
I)
,
a
nd
MA,
and
ti
me
serie
s
do
not
requir
e
an
integ
rated
par
t
to
decli
ne
the
seaso
nalit
y
re
pr
ese
nted
as
mixed
aut
or
e
gressi
ve
a
nd
m
ov
i
ng
ave
rage
(A
R
MA)
models,
al
l
desi
gned
to
captu
re th
e
fu
ndame
ntal patt
e
rn
s
and t
rends
i
nh
e
re
nt in
ti
me
ser
ie
s
data
[40]
.
a.
Au
t
or
e
gr
e
ssive
(
AR
)
,
t
he
e
qu
at
ion
for AR
of o
rd
e
r
p
ca
n be
w
ritt
en
as
(7)
:
=
+
∅
1
−
1
+
∅
2
−
2
+
⋯
+
∅
−
+
(7)
b.
M
ovi
ng
a
ve
rage (
M
A
)
,
t
he
e
quat
ion f
or
M
A
of ord
e
r q ca
n be
wr
it
te
n
as
(
8)
:
=
−
1
−
1
−
2
−
2
−
⋯
−
−
+
(8)
c.
M
ixe
d
a
utoreg
ressive
and
movin
g
a
ver
a
ge
(
ARMA)
,
t
he
e
qu
at
io
n f
or
AR
M
A
(p,
q) ca
n be
wr
it
te
n
as
(
9)
:
=
+
∅
1
−
1
+
∅
2
−
2
+
⋯
+
∅
−
−
1
−
1
−
2
−
2
−
⋯
−
−
+
(9)
d.
Au
t
or
e
gr
e
ssive
integrate
d
m
ovin
g
a
ver
a
ge
(
ARI
M
A
)
,
t
he
equ
at
io
n
for
A
RIMA
(
,
,
)
ca
n
be
wr
it
te
n
as
(10)
:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
728
-
740
734
=
−
1
+
∅
1
−
1
−
∅
2
−
2
+
−
1
−
1
(10)
wh
e
re
is
the
act
ual
value
a
nd
is
the
ra
ndom
error.
is
the
mean
a
bout
whic
h
the
se
ries
fl
uctuates.
The
n,
∅
is
pa
ramete
rs
of
aut
or
e
gr
e
ssive
a
nd
is
the
movin
g
a
ver
a
ge
pa
rameters
t
o
be
est
imat
e
d.
−
is
the
e
rro
r
te
rms
(
=
1
,
2
,
3
…
)
a
ssumed
to
be
in
de
pende
ntly
dis
tribu
te
d
over
t
ime.
ℎ
order
of
the
la
gge
d
la
gg
e
d
dep
e
ndent
or c
urren
t
value
.
3.2.
Pr
oposed
m
eth
od
The
f
ormulat
ion
of
m
od
el
de
velo
pm
e
nt
for
the
HIM
A
i
nvolv
e
s
th
e
tr
aditi
on
al
H
olt
method
an
d
movin
g
a
ver
a
ge
f
r
om B
ox
-
Je
nk
i
ns
meth
odol
ogy.
3.2.1
Movin
g
a
ver
age
(MA) fr
om
B
ox
-
Je
n
kins
m
et
hodol
og
y
The
MA
m
ode
l
in
the
Bo
x
-
Je
nk
i
ns
meth
odol
ogy
is
a
pa
st
f
or
ecast
er
r
or
(
mu
lt
ipli
ed
by
a
coe
ff
ic
ie
nt)
in a
regressio
n
-
li
ke
m
odel
.
M
A:
=
−
1
−
1
−
2
−
2
−
⋯
−
−
+
=
(
−
∑
−
=
1
+
)
(1
1
)
wh
e
re
is
t
he
a
ct
ual
val
ue
a
nd
is
the
ra
ndom
er
ror,
ass
um
e
d
t
o
be
i
den
ti
c
al
ly,
in
de
pend
ently
distrib
ute
d,
with a
mean
e
qual
to
ze
r
o
a
nd the sa
me
var
ia
nce.
3.2.2. Hol
t
int
egrated
movin
g av
er
age
(HIMA
)
In
reali
ty,
ti
me
series
data
co
ns
ist
s
of
li
nea
r
an
d
non
-
li
nea
r
com
pone
nts
t
hat
ma
ke
up
t
he
moveme
nt
tren
ds
[
41]
,
[
42]
.
H
I
M
A
m
odel
is
the
c
ombinati
on
of
t
he
Ho
lt
met
hod
an
d
t
he
MA
f
r
om
B
ox
-
J
enk
i
ns
method
ology.
The
pro
posed
new
hybr
i
d
m
od
el
is
dev
el
oped
to
e
nh
a
nc
e
f
or
ecast
accu
racy
by
inc
orp
or
at
in
g
the
resi
dual
va
lues
f
r
om
H
olt
meth
od
int
o
a
MA
model.
A
new
strat
e
gy
is
i
ntr
oduce
d
wh
e
re
t
he
re
sidu
al
value
in
(12)
i
s
resu
lt
in
g
f
rom
the
est
imat
ed
resi
du
al
val
ue
wer
e
ob
ta
in
from
tra
diti
on
a
l
Ho
lt
meth
od
that
us
e
d
as a
n
i
nput
v
aria
ble to t
he
MA
model.
Ho
lt
resid
ual
:
=
−
̂
(12)
HIM
A
m
odel
:
∗
+
=
(
+
×
)
±
(
−
∑
−
=
1
+
)
(13)
wh
e
re
as
error
values
,
w
hich
are
the d
iffe
re
nc
e
betwee
n
act
ual
data
and
t
he
fitt
ed
value
s
̂
and
+
∗
is
the
f
or
eca
st
va
lue
that
will
be
ge
ner
at
e
d
f
r
om
t
he
a
dd
it
io
n
or
s
ubtract
io
n
of
t
he
f
orec
ast
values
ge
ne
rated
from
t
rad
it
io
na
l Ho
lt
’s
meth
od.
Ther
e
are
th
re
e
reas
ons
for
com
bin
in
g
wit
h
MA:
i
)
e
rro
r
co
rr
ect
io
n,
w
hich
mea
ns
MA
wor
ks
t
o
correct
an
y
e
rror
s
pro
duced
by
the
AR
mod
el
in
pro
duci
ng
f
oreca
st
val
ue
s.
F
or
t
he
HIMA
m
od
el
,
th
e
MA
model
will
fi
x
an
y
e
rror
s
ma
de
by
H
olt
me
thod
;
ii
)
i
m
pro
ve
acc
ur
ac
y
pr
edict
ion
by
ta
ki
ng
past
e
rro
r
t
erms
into
acc
ount
,
Ho
lt
meth
od
c
an
rea
dju
st
it
s
pr
e
dicti
on
s
acc
ordin
gly
to
im
pro
ve
t
he
valu
e
of
f
or
ecast
a
ccur
ac
y
;
and
iii
)
r
e
duce
resid
uals
or
di
ff
e
ren
ces
betw
een
pre
dicte
d
a
nd
act
ual
valu
es.
This
le
a
ds
to
m
or
e
e
ff
ic
ie
nt
an
d
eff
ect
ive
m
od
e
ls for
pr
e
dicti
ng
fu
t
ur
e
value
s
[
43]
.
3.3.
M
od
el
e
valua
tio
n
On
ce
the
m
od
el
is
dev
el
op
e
d,
it
is
eval
ua
te
d
us
i
ng
the
te
sti
ng
data
.
This
e
valuati
on
in
volves
app
l
ying
e
rror met
rics
to
asse
ss
the
m
od
el
's
accur
ac
y.
Er
r
or
mea
sures
a
re u
se
d
to
di
ff
e
re
ntiat
e
betwee
n a
po
or
and
a
good
f
or
ecast
model.
I
n
ot
her
wor
ds
,
the
erro
r
meas
ur
e
was
us
ed
t
o
fin
d
the
best
model
w
hich
fits
the
data.
A
m
odel
that
has
the
s
mall
est
error
i
s
sai
d
t
o
be
th
e
best
m
odel
.
The
e
valuati
on
measu
res
cal
c
ulate
d
include
the
r
oo
t
mean
squa
re
error
(R
M
SE
),
mean
abs
olu
te
error
(
MAE)
,
a
nd
mea
n
a
bsolute
pe
rc
e
ntage
error
(MAPE
)
val
ue
s
f
or
eac
h
m
odel
.
T
he
f
or
m
ul
as
us
ed
to
cal
c
ulate
the
er
r
or
s
are:
=
−
̂
,
w
he
re
i
s
the
act
ual
value
i
n
ti
me
(
t
)
an
d
̂
re
fer
s
t
o
fitt
ed
va
lue
in
ti
me
(
t
).
Roo
t
mea
n
s
qu
are
er
ror
(RM
S
E)
is
a
sta
ti
sti
cal
metri
c
em
ploy
ed
to
gauge
t
he
de
gr
ee
of
e
rror
in
predict
io
ns
by
c
ompu
ti
ng
the
squa
re
r
oo
t
of
the
a
verage
of
the s
qu
a
red di
f
fer
e
nces
betwe
en pre
dicte
d
a
nd actual
val
ues
.
RMSE =
√
∑
2
(14)
M
ea
n
a
bsolute
er
ror
(MAE
)
is
an
oth
e
r
met
ric
use
d
f
or
er
ror
asse
ssme
nt
,
f
oc
us
in
g
on
t
he
a
ver
a
ge
abs
olu
te
diff
e
re
nces
be
tween
predict
ed
a
nd
act
ual
value
s,
maki
ng
it
le
ss
in
fluen
ce
d
by
e
xt
reme
ou
tl
ie
rs
w
hen
com
par
e
d
t
o
R
M
SE
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
En
hancin
g
P
E
TRO
NA
S
s
hare
p
ric
e f
or
eca
sts:
a hyb
ri
d
H
olt i
ntegr
ated
…
(
Nuri
n
Qisti
na
Mo
hamad Fo
zi
)
735
M
A
E =
∑
|
|
(15)
M
ea
n
ab
so
l
ute
per
ce
nta
ge
er
r
or
(
M
A
PE)
pr
ov
i
des
a
perce
ntage
-
base
d
as
sessme
nt
of
er
r
or
betwee
n
pre
dicte
d
and
act
ual
val
ues.
It
pro
ves
par
ti
cula
rly
valuab
le
wh
e
n
e
xpressi
ng
er
ror
as
a
pro
portio
n
of
the
act
ual
values
,
facil
it
at
ing
compa
risons ac
ross d
i
ver
se
d
at
a
set
s and m
odel
s.
Wh
e
re
n
is s
ample size
.
M
A
PE =
∑
(
)
∗
1
00
(16)
4.
RESU
LT
S
AND DI
SCUS
S
ION
Figure
4
il
lust
r
at
es
the
or
i
gin
a
l
ti
me
series
pl
ot
da
il
y
P
ETR
ONAS
f
or
hi
gh
s
har
e
pr
ic
e
i
n
Fi
gure
4(a)
and
lo
w
s
hare
pr
ic
e
i
n
Fi
gure
4(b
)
.
The
ti
me
series
plo
t
f
or
daily
P
ETRO
NAS
high
s
hare
pri
ce
in
Fig
ure
4
(a
)
and
lo
w
s
har
e
pr
ic
e
in
Fig
ure
4
(
b)
f
r
om
Jan
uary
3,
2000
to
Octo
be
r
2,
2023,
rev
eal
s
upwa
r
d
a
nd
do
wnward
tren
d
moveme
nt.
I
niti
al
ly,
th
e
s
har
e
pri
ce
e
xp
e
rience
d
a
de
cl
ine
un
ti
l
a
ppr
oximat
el
y
Ja
nu
a
r
y
1,
2004,
the
n
after
wh
ic
h
it
sta
bili
zed
with
minimal
fl
uctuati
ons.
A
sig
nificant
upwa
r
d
tre
nd
is
obse
rv
e
d
from
Ja
nuar
y
3,
2007
to
Dece
mb
e
r
31,
20
13,
culminati
ng
i
n
a
pea
k
a
rou
nd
2014
to
20
15.
Fo
ll
owin
g
this
peak,
the
sh
a
re
pr
ic
e
decli
ned
a
nd
then
ente
red
a
ph
a
se
of
sta
bili
zat
ion
,
w
hich
be
gan
ar
ound
Jan
uary
4,
2016.
F
r
om
Jan
ua
ry
3
,
2020
t
o
Oct
ober 2
,
2
02
3,
the
sh
are
p
rice
re
mained
r
el
at
iv
el
y
sta
ble. Th
is
init
ia
l
anal
ys
is
s
uggests
that
t
he
H
olt
method is
ef
fe
ct
ive in ca
pturing t
he
t
rend
pa
tt
ern
s
pr
ese
nt i
n
the
d
at
a.
(a)
(b)
Figure
4. Time
series
plo
t
for dai
ly
(a)
h
ig
h a
nd (b)
l
ow
4.1.
M
od
el
p
e
rfo
rm
an
ce
The
final
ste
p
is
to
i
nter
pr
et
the
m
odel
's
pe
rformance
,
hig
hlig
htin
g
it
s
eff
ect
ive
ness
in
reducin
g
resid
uals
a
nd
impro
ving
pr
e
dicti
on
accu
ra
cy.
Table
2
s
hows
the
e
rro
r
mea
sures
a
nd
forecast
acc
ur
ac
y
per
ce
ntage
f
or
daily
hi
ghs.
The
[
Ho
lt
+
M
A
]
an
d
HIM
A
[
Ho
lt
–
M
A
]
mod
el
s
present
the
best
pe
rforma
nces
base
d
on
t
he
lowest
e
rror
me
asur
e
s
a
nd
t
he
highest
f
or
ecas
t
accu
racy
per
c
entage.
MA
m
od
el
or
va
lue
of
q
i
s
the
num
ber
of
sig
nificant
s
pik
es
in
t
he
autoc
orrelat
ion
functi
on
(
A
CF).
T
he
be
st
M
A
model
f
or
the
[Holt
+M
A]
bet
ween
set
1
to
s
et
5
is
M
A
(
1)
,
M
A
(
4)
,
M
A
(
4),
M
A
(
3)
,
a
nd
M
A
(
1)
.
M
ea
nwhile
,
MA
(1)
for
set
1
t
o
set
4
an
d MA(
6)
f
or
s
et
5
h
a
d
the
be
st MA model.
T
he
f
oreca
st
acc
uracy
p
erce
ntag
e
ra
nged
f
r
om 56.76%
to
99.
68
%
,
de
monstrati
ng
sign
i
ficant
va
riabil
it
y
in
pr
e
di
ct
i
ve
performa
nce
am
ong
th
e
m
odel
s.
The
be
s
t
model
pe
rform
ance
f
or
each
model
(
Ho
lt
method,
DT
M,
ARI
MA,
HIMA
[
Ho
lt
+
MA],
an
d
HIM
A
[
H
olt
–
M
A
]) is set
1 wit
h
the
train
-
t
est
sp
li
t o
f 99
:
1.
Table
3
s
hows
the
er
r
or
meas
ur
es
an
d
perce
ntage
of
f
oreca
st
accurac
y
f
or
da
il
y
l
ow
s
.
T
he
best
M
A
model
f
or
the
[Holt
+M
A]
is
M
A
(
2)
,
MA(
1),
M
A
(3),
MA
(5),
a
nd
MA(
4)
f
or
set
1
to
set
5,
res
pecti
vely.
M
ea
nwhile
, the
H
I
MA [Ho
lt
–
MA
]
m
od
el
s
hows
t
hat all
sets of
MA(
1)
a
re th
e
best MA
models, e
xclu
ding set
2,
wh
ic
h
is
M
A
(
7)
.
The
f
or
ecast
acc
ur
a
cy
pe
rce
ntage
ranged
from
56.99%
to
98.
88%,
de
m
onstrat
in
g
sign
ific
a
nt
var
i
abili
ty
in
pre
dicti
ve
pe
rfo
rma
nce
a
mon
g
t
he
m
od
el
s
.
F
or
t
he
daily
lo
ws,
the
best
perf
ormance
for
eac
h
m
odel
(
H
olt
meth
od,
D
TM,
ARI
MA,
H
I
M
A
[Hol
t+
M
A
],
an
d
H
I
M
A
[Holt
–
MA
])
is
al
s
o
set
1
with
the
trai
n
-
te
st
s
plit
of
99:1.
F
or
t
he
HIM
A
m
od
el
,
bo
t
h
dail
y
high
an
d
lo
w
do
not
re
qu
ir
e
di
ff
e
ren
ci
ng
be
cause
the p
-
values
a
r
e less t
ha
n 0.0
5.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
omp E
ng,
V
ol.
15
, No
.
1
,
Febr
uary
20
25
:
728
-
740
736
Table
2
.
E
rro
r mea
sures a
nd
per
ce
ntage
of
f
or
ecast
ac
cu
rac
y for eac
h
m
od
el
(
daily
h
i
gh
)
Mod
el
Set
MA
(
q
)
T
rainin
g
:
t
estin
g
RMSE
(
testin
g
)
MAE
(
testin
g
)
MAPE
(
testin
g
)
Percentag
e of
f
o
re
cast
a
ccuracy
(
t
estin
g
)
Ho
lt’
s
m
eth
o
d
1
-
9
9
:1
0
.20
6
4
0
.18
0
5
1
.04
6
5
9
9
.68
2
-
9
5
:5
0
.52
5
9
0
.45
6
0
2
.65
8
2
9
7
.34
3
-
9
0
:1
0
0
.70
3
6
0
.60
9
1
3
.57
2
1
9
6
.43
4
-
8
0
:2
0
3
.54
1
6
3
.37
2
1
2
0
.20
6
0
7
9
.79
5
-
7
0
:3
0
7
.98
3
4
7
.37
2
6
4
3
.24
0
6
5
6
.76
DTM
1
-
9
9
:1
0
.24
3
8
0
.21
6
3
1
.25
4
0
9
9
.59
2
-
9
5
:5
0
.67
7
5
0
.62
5
8
3
.65
4
4
9
6
.35
3
-
9
0
:1
0
1
.04
1
6
0
.95
3
1
5
.59
4
4
9
4
.41
4
-
8
0
:2
0
2
.24
7
7
2
.13
1
0
1
2
.82
2
2
8
7
.18
5
-
7
0
:3
0
4
.69
4
2
4
.41
6
7
2
5
.86
0
6
7
4
.14
ARIMA
1
-
9
9
:1
0
.24
9
2
0
.22
1
9
1
.28
7
0
9
9
.58
2
-
9
5
:5
0
.66
3
4
0
.61
0
8
3
.56
6
7
9
6
.43
3
-
9
0
:1
0
1
.03
1
7
0
.94
3
7
5
.53
9
1
9
4
.46
4
-
8
0
:2
0
2
.21
7
0
2
.09
9
7
1
2
.63
5
7
8
7
.36
5
-
7
0
:3
0
7
.86
2
2
7
.26
3
5
4
2
.60
0
8
5
7
.40
HIM
A
[Holt
+
MA]
1
MA(
1
)
9
9
:1
0
.20
6
5
0
.18
0
6
1
.04
7
3
9
9
.68
2
MA(
4
)
9
5
:5
0
.52
5
9
0
.45
5
9
2
.65
8
0
9
7
.34
3
MA(
4
)
9
0
:1
0
0
.70
3
7
0
.60
9
3
3
.57
3
0
9
6
.43
4
MA(
3
)
8
0
:2
0
3
.54
1
8
3
.37
2
3
2
0
.20
7
4
7
9
.79
5
MA(
1)
7
0
:3
0
7
.98
4
1
7
.37
3
3
4
3
.24
4
5
5
6
.76
HIM
A
[Holt
−
MA]
1
MA(
1
)
9
9
:1
0
.20
6
3
0
.18
0
3
1
.04
5
7
9
9
.68
2
MA(
1
)
9
5
:5
0
.52
6
0
0
.45
6
0
2
.65
8
3
9
7
.34
3
MA(
1
)
9
0
:1
0
0
.70
3
4
0
.60
8
9
3
.57
1
0
9
6
.43
4
MA(
1
)
8
0
:2
0
3
.54
1
3
3
.37
1
8
2
0
.20
4
5
7
9
.80
5
MA(
6
)
7
0
:3
0
7
.98
2
7
7
.37
1
9
4
3
.23
6
6
5
6
.76
Table
3
. E
rro
r mea
sures a
nd
per
ce
ntage
of
f
or
ecast
ac
cu
rac
y for eac
h
m
od
el
(
daily
lo
w)
Mod
el
Set
MA
(
q
)
T
rainin
g
:
testin
g
RMSE
(testin
g
)
MAE
(testin
g
)
MAPE
(testin
g
)
Percentag
e of
f
o
re
cast
accuracy
(
t
estin
g
)
Ho
lt’
s
m
eth
o
d
1
-
9
9
:1
0
.20
6
1
0
.19
0
6
1
.1
1
8
9
9
8
.88
2
-
9
5
:5
0
.39
7
8
0
.33
2
0
1
.96
5
1
9
8
.04
3
-
9
0
:1
0
0
.61
6
2
0
.52
8
6
3
.14
9
6
9
6
.85
4
-
8
0
:2
0
3
.14
0
9
2
.98
0
6
1
8
.20
3
1
8
1
.80
5
-
7
0
:3
0
7
.68
0
5
7
.1
1
4
3
4
2
.50
4
0
5
7
.50
DTM
1
-
9
9
:1
0
.24
4
3
0
.22
3
9
1
.31
4
0
9
8
.69
2
-
9
5
:5
0
.56
2
2
0
.50
4
1
2
.98
4
8
9
7
.02
3
-
9
0
:1
0
0
.96
7
2
0
.88
0
8
5
.25
0
6
9
4
.75
4
-
8
0
:2
0
1
.93
8
1
1
.81
2
7
1
1
.13
4
4
8
8
.87
5
-
7
0
:3
0
4
.71
1
5
4
.43
6
5
2
6
.47
1
2
7
3
.53
ARIMA
1
-
9
9
:1
0
.24
3
6
0
.22
3
3
1
.31
0
4
9
8
.69
2
-
9
5
:5
0
.55
8
6
0
.50
0
4
2
.96
2
3
9
7
.04
3
-
9
0
:1
0
0
.96
1
4
0
.87
5
4
5
.21
8
0
9
4
.78
4
-
8
0
:2
0
1
.96
3
9
1
.83
9
2
1
1
.29
5
9
8
8
.70
5
-
7
0
:3
0
7
.78
2
3
7
.19
8
4
4
3
.01
3
5
5
6
.99
HIM
A
[Holt
+
MA]
1
MA(
2
)
9
9
:1
0
.20
6
2
0
.19
0
6
1
.1
1
9
1
9
8
.88
2
MA(
1
)
9
5
:5
0
.39
8
0
0
.33
2
3
1
.96
6
3
9
8
.03
3
MA(
3
)
9
0
:1
0
0
.61
6
4
0
.52
8
7
3
.15
0
2
9
6
.85
4
MA(
5
)
8
0
:2
0
3
.14
1
0
2
.98
0
7
1
8
.20
3
7
8
1
.80
5
MA(
4
)
7
0
:3
0
7
.68
1
0
7
.1
1
4
9
4
2
.50
7
5
5
7
.49
HIM
A
[Holt
−
MA]
1
MA(
1
)
9
9
:1
0
.20
6
0
0
.19
0
5
1
.1
1
8
2
9
8
.88
2
MA(
7
)
9
5
:5
0
.39
8
0
0
.33
2
3
1
.96
6
4
9
8
.04
3
MA(
1
)
9
0
:1
0
0
.61
6
1
0
.52
8
5
3
.14
8
9
9
6
.85
4
MA(
1
)
8
0
:2
0
3
.14
0
7
2
.98
0
5
1
8
.20
2
4
8
1
.80
5
MA(
1
)
7
0
:3
0
7
.67
9
9
7
.1
1
3
7
4
2
.50
0
5
5
7
.50
In
summa
ry,
al
l
models
ac
hie
ve
t
heir
best
pe
rformance
wit
h
a
99:1
trai
ni
ng/t
est
ing
rati
o.
Among
the
models,
the
tr
aditi
on
al
H
olt
method
a
nd
t
he
propose
d
H
I
M
A
[Holt
+
M
A]
model
dem
on
strat
e
the
high
est
accurac
y
a
nd
l
ow
est
er
ror
r
at
es,
pa
rtic
ularl
y
in
the
99:1
sc
enar
i
o.
T
he
pe
rformance
of
a
ll
mo
dels
dete
r
iorates
as
the
pro
port
ion
of
trai
ni
ng
data
decr
ea
se
s,
em
phasi
zi
ng
the
im
portan
ce
of
a
la
rg
e
r
trai
ning
datas
et
for
accur
at
e
f
or
ec
ast
ing
.
The
previo
us
st
udy
by
Ce
r
quei
ra
et
al.
[
44]
in
dic
at
es
that
the
siz
e
of
the
trai
ni
ng
set
gr
eat
ly
impact
s
the
acc
ur
ac
y
of
forecast
i
ng
m
od
el
s.
L
arg
e
r
trai
ning
set
s
te
nd
to
impro
ve
performa
nce,
especial
ly
wh
e
n
deali
ng
with
data
that
is
not
sta
ti
on
a
ry.
Table
s
4
a
nd
5
disp
la
y
the
model
co
mp
a
ri
so
n
f
or
eac
h
m
odel
with
the
best
set
of
data
pa
rtit
ion
in
g
for
daily
high
a
nd
lo
w,
resp
ect
ivel
y.
T
his
te
sti
ng
ph
a
se
is
desig
ne
d
to
ri
gor
ously
e
val
ua
te
the
m
odel
's
performa
nce
on
a
substa
ntial
te
sti
ng
set
,
en
s
ur
i
ng
the
co
mpa
rison
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
omp E
ng
IS
S
N:
20
88
-
8708
En
hancin
g
P
E
TRO
NA
S
s
hare
p
ric
e f
or
eca
sts:
a hyb
ri
d
H
olt i
ntegr
ated
…
(
Nuri
n
Qisti
na
Mo
hamad Fo
zi
)
737
model
can
ge
ne
rali
ze
well
to n
ew
data,
hel
pin
g
to max
imi
z
e
predict
ive
ac
cur
ac
y
i
n
pr
e
dicti
ng
s
har
e
pri
ces
f
or
diff
e
re
nt p
e
rio
ds
.
Table
4
. M
od
el
co
m
pa
rison
for
eac
h
m
odel
wi
th the best
set
of d
at
a
par
ti
ti
onin
g (d
ai
ly
h
i
gh)
Mod
el
Set
T
rainin
g
:
testin
g
RMSE
(testin
g
)
MAE
(testin
g
)
MAPE
(testin
g
)
Percentag
e of
f
o
re
cast
accuracy
(
t
estin
g
)
Ho
lt’
s
m
eth
o
d
1
9
9
:1
0
.20
6
4
0
.18
0
5
1
.04
6
5
9
9
.68
Dam
p
ed
t
r
en
d
m
et
h
o
d
1
9
9
:1
0
.24
3
8
0
.21
6
3
1
.25
4
0
9
9
.59
ARIMA
1
9
9
:1
0
.24
9
2
0
.22
1
9
1
.28
7
0
9
9
.58
HIM
A
[Holt
+
MA]
1
9
9
:1
0
.20
6
5
0
.18
0
6
1
.04
7
3
9
9
.68
HIM
A
[Holt
–
MA]
1
9
9
:1
0
.20
6
3
0
.18
0
3
1
.04
5
7
9
9
.68
Table
5
. M
od
el
co
m
pa
rison
for
eac
h
m
odel
wi
th the best
set
of d
at
a
par
ti
ti
onin
g (d
ai
ly
lo
w
)
Mod
el
Set
T
rainin
g
:
testin
g
RMSE
(testin
g
)
MAE
(testin
g
)
MAPE
(testin
g
)
Percentag
e of
f
o
re
cast
accuracy
(
t
estin
g
)
Ho
lt’
s
Metho
d
1
9
9
:1
0
.20
6
1
0
.19
0
6
1
.1
1
8
9
9
8
.88
Dam
p
ed
T
rend
M
e
th
o
d
1
9
9
:1
0
.24
4
3
0
.22
3
9
1
.31
4
0
9
8
.69
ARIMA
1
9
9
:1
0
.24
3
6
0
.22
3
3
1
.31
0
4
9
8
.69
HIM
A
[Holt +
M
A
]
1
9
9
:1
0
.20
6
2
0
.19
0
6
1
.1
1
9
1
9
8
.88
HIM
A
[Holt
–
MA
]
1
9
9
:1
0
.20
6
0
0
.19
0
5
1
.1
1
8
2
9
8
.88
The
re
su
lt
of
this
study
s
ho
ws
that
pro
posed
HIM
A
m
od
el
s
ha
d
lo
w
er
er
ror
rates
and
highest
accurac
y
tha
n
DT
M
,
AR
IMA
m
od
el
s
,
a
nd
t
r
aditi
on
al
H
olt
methods
f
or
bo
th
daily
hi
gh
a
nd
lo
w
PETR
ONAS
sh
are
pr
ic
es.
Althou
gh
tra
di
ti
on
al
H
olt
a
nd
pr
opos
e
d
HIM
A
m
odel
s
sho
w
simi
la
r
accu
racy
rate
s,
the
pro
po
se
d
H
IMA
[Holt
–
MA]
model
out
perf
orms
oth
e
r
f
or
ecast
ing
meth
od
s
.
T
his
res
ul
t
is
con
sist
e
nt
with
wh
at
has
bee
n
repo
rted
by
Ha
nsun
a
nd
S
ub
a
na
r
[45]
in
their
w
or
k
on
a
new
hybri
d
m
od
el
cal
le
d
hybri
d
-
weig
hte
d
e
xpone
ntial
movin
g
ave
rage
(
H
-
WE
MA
)
,
w
hich
mer
ges
the
cal
culat
io
n
of
t
he
wei
ghti
ng
factor
i
n
wei
ghte
d
MA w
it
h
t
he
H
olt met
ho
d
f
or
Ja
ka
rta
stock
e
xc
hange
com
posit
e ind
e
x.
T
he
c
urren
t
fin
ding
sh
ows
t
hat
the
hybri
d
H
-
WE
M
A
pro
vide
s
more
accu
rate
an
d
r
obus
t
forecast
ing
res
ul
ts
com
par
e
d
t
o
the
tr
aditi
on
al
weigh
te
d
m
ov
i
ng
aver
a
ge
(
W
MA)
a
nd
H
olt
m
et
hod.
T
his
res
ult
furthe
r
rei
nfo
rces
m
os
t
fi
nd
i
ngs
that
hy
br
id
m
odel
s
ca
pture
c
omplex
patte
r
ns
i
n
ti
me
s
eri
es
data,
le
adi
ng
to
m
ore
acc
ur
at
e
a
nd
de
pe
nd
a
ble
forecast
s
[46]
.
The
stu
dy
sho
wcases
a
hybr
i
d
ap
proac
h
tha
t
le
ver
ages
t
he
stren
gth
s
of
t
he
tradit
io
nal
Ho
lt
meth
od
and
the
MA
model,
yieldin
g
im
pro
ved
pr
edict
ive
pe
rfo
r
mance,
pa
rtic
ul
arly
f
or
PE
T
RON
AS
s
ha
re
pr
ic
e
dataset
.
T
his
c
ombinati
on
ha
r
nesses
the
H
olt
meth
od's
ca
pa
bili
ty
to
ca
pt
ure
li
near
tren
ds
and
the
M
A
m
od
el
's
abili
ty
to
s
moo
th
out
s
hort
-
te
r
m
fl
uctuati
ons,
res
ulti
ng
i
n
m
or
e
accu
rate
an
d
reli
a
ble
f
or
ec
ast
s.
By
i
nteg
r
at
in
g
these
te
ch
niqu
es,
the
hybri
d
mod
el
a
ddres
ses
the
li
mit
at
ion
s
i
nh
e
re
nt
in
us
in
g
ei
the
r
meth
od
al
one
,
th
us
pro
vid
in
g
a
m
or
e
r
obus
t
f
oreca
sti
ng
t
oo
l
f
or
fin
ancial
ti
me
se
ries
data.
T
he
a
ppli
cat
i
on
of
t
his
m
odel
t
o
PETRO
NAS
s
har
e
pr
ic
e
hi
ghli
gh
ts
it
s
pr
a
ct
ic
al
releva
nc
e
an
d
pote
ntial
f
or
en
h
a
ncin
g
decisi
on
-
ma
king
processes
in fi
nan
ci
al
a
nalysi
s and i
nv
est
me
nt strategies
.
Con
tra
r
y
to
w
ha
t
has
bee
n
re
porte
d
by
Ai
rlan
gg
a
et
al.
[
47]
,
the
res
ults
indi
cat
e
the
backp
ropa
gation
neural
net
wor
ks
al
gorith
m
outpe
rforms
co
mp
a
red
t
o
the
tradit
ion
al
sin
gle,
do
ub
le
,
a
nd
tri
ple
ex
po
nen
ti
al
smoothi
ng
m
odel
s
in
te
rms
of
accu
rac
y,
achievi
ng
lo
wer
e
rro
r
rate
s
f
or
rice
pro
du
ct
io
n
i
n
I
ndones
ia
.
Ther
e
f
or
e,
f
or
fu
t
ur
e
resea
r
ch,
t
his
stu
dy
sugg
e
sts
the
app
li
cat
io
n
of
oth
e
r
hybr
i
d
models
that
c
ombine
var
i
ou
s
mac
hi
ne
-
le
ar
ning
t
echn
i
qu
e
s
wit
h
tra
diti
on
al
sta
ti
sti
cal
methods
.
F
or
instance,
in
te
gr
at
in
g
backp
ropa
gation
neural
netw
orks
with
met
hods
li
k
e
AR
I
MA,
H
olt
-
Wi
nte
rs,
or
e
xpone
ntial
smoothi
ng
cou
l
d
yield
e
ve
n
mor
e
accu
rate
a
nd
rob
us
t
f
oreca
st
ing
m
od
el
s
.
A
dd
it
io
nally,
a
dvance
d
mac
hine
le
ar
ning
al
gorithms
su
c
h
as
s
uppo
rt
vect
or
mac
hi
nes
(SVM),
r
andom
f
orest
s
,
or
Gr
a
dient
Boo
sti
ng
co
ul
d
be
i
ncor
pora
t
ed
to
captu
re c
omplex
non
-
li
near
pa
tt
ern
s a
nd inter
act
ion
s
within
the d
at
a.
Fu
tu
re
resea
rc
h
sho
uld
f
oc
us
on
r
efini
ng
t
he
se
hybri
d
m
odel
s
by
opti
mizi
ng
their
para
mete
rs
an
d
impro
ving
th
e
ir
co
mputat
io
nal
ef
fici
enc
y.
Co
mp
a
rati
ve
stu
dies
c
ou
l
d
be
c
onduct
ed
t
o
e
valuat
e
the
performa
nce
of
diff
e
ren
t
hy
br
i
d
m
odel
s
a
cro
ss
va
rio
us
dataset
s,
in
cl
udin
g
th
os
e
wit
h
diff
e
re
nt
te
mporal
reso
l
utions
(d
a
il
y,
week
l
y,
m
on
t
hly)
a
nd
c
ha
racteri
sti
cs
(li
near
a
nd
no
n
-
li
near
tre
nds).
M
ore
ov
e
r,
it
w
ou
l
d
be
ben
e
fici
al
to
e
xp
l
or
e
the
ap
pl
ic
abili
ty
of
t
he
se
m
od
el
s
to
ot
her
fi
nan
ci
al
m
ark
et
s,
c
om
m
odit
ie
s,
a
nd
ec
onomi
c
ind
ic
at
ors
t
o
validat
e
thei
r
ef
fecti
ven
e
ss
an
d
ge
ner
al
i
zabil
it
y
un
der
di
ver
se
co
nd
it
ion
s.
Furthe
r
more,
inco
rpor
at
in
g
domain
-
s
pec
ific
knowle
dg
e
an
d
e
xter
na
l
factors,
s
uc
h
as
m
acr
oe
conomic
i
ndic
at
or
s,
geop
olit
ic
al
even
ts,
or
in
dust
ry
-
spe
ci
fic
vari
ables,
c
ou
l
d
e
nh
a
nce
the
f
oreca
sti
ng
ac
cu
r
acy
of
these
hybri
d
models.
I
nv
est
igati
ng
the
inte
gr
at
io
n
of
real
-
ti
me
data
st
rea
ms
a
nd
a
dap
ti
ve
le
arn
i
ng
mec
han
is
ms
w
ou
l
d
al
so
be valua
ble for
de
velo
ping m
od
el
s t
hat ca
n dynamica
ll
y
a
dj
us
t t
o
c
hangi
ng ma
rket
cond
i
ti
on
s. Ult
imat
el
y,
the
Evaluation Warning : The document was created with Spire.PDF for Python.